USING NONLINEAR MODELS TO DESCRIBE HEIGHT
GROWTH CURVES IN PANTANEIRO HORSES
1SANDRA APARECIDA SANTOS2, GERALDO DA SILVA E SOUZA3, MARCOS RUBEN DE OLIVEIRA4 and JOSÉ ROBSON SERENO5
ABSTRACT - Height at withers data mostly from birth to 36 months of age of 26 Pantaneiro horses were used to fit Brody, Richards, Gompertz, Logistic, Weibull and Morgan-Mercer-Flodin nonlinear response functions. Based on measures of average curvature and combined mean square error, the Weibull model was chosen. The asymptote of this curve, representing the average height at maturity, was higher for males than females. The maturity index, however, was more elevated for females than males. There was indication of a negative association between the maturity index and height at matu-rity only for females. These results indicate that females mature earlier. After checking for normality and homogeneity of variances within groups (sex) the analysis investigated sex differences via t-tests. A significant difference was detected only for height at birth.
Index terms: curvature, growth, height-age relationships.
USO DE MODELOS NÃO-LINEARES PARA O AJUSTE DE CURVAS DE CRESCIMENTO DE CAVALOS PANTANEIROS
RESUMO - Dados de altura da cernelha e idade de 26 cavalos Pantaneiros, obtidos, na maioria, do nascimento até 36 meses, foram ajustados aos modelos de respostas não-lineares de Brody, Richards, Gompertz, Logístico, Weibull e Morgan-Mercer-Flodin. Estes seis modelos matemáticos foram com-parados com o uso de uma medida de curvatura média e do erro médio quadrático combinado. O modelo de Weibull foi escolhido. A assíntota desta curva representa a altura esperada na maturidade. Os machos apresentaram um valor maior desta quantidade do que as fêmeas. O índice de maturidade, contudo, é maior para as fêmeas. Observou-se uma indicação de associação negativa entre altura na maturidade e índice de maturidade somente nas fêmeas. Tais resultados indicam que as fêmeas amadu-recem mais cedo. Após testes de normalidade e homogeneidade de variância, as diferenças entre sexos foram analisadas com o uso do teste-t. Somente a altura da cernelha ao nascimento apresentou diferen-ça significativa.
Termos para indexação: curvatura, crescimento, relações altura-idade.
(Robinson, 1976). Hyperplasy (an increase in cell
number) and hypertrophy (an increase in cell size)
determine to what extent an animal increases in
weight and size, and development determines the
shape of an animal along with various organ
func-tions. The study of animal growth is often described
as the integration of all aspects of animal science
including nutrition, genetics or animal breeding,
physiology and meat science (Trenkle & Marple,
1983).
Growth curves reflect the lifetime
interrelation-ships between an individuals inherent impulse to
grow and mature all body parts and the environment
in which these impulses are expressed. Knowledge
of the growth curve is important to all animal
scien-1Accepted for publication on February 18, 1999.
2Zootecnist, M.Sc., Embrapa-Centro de Pesquisa Agropecuária
do Pantanal (CPAP), Caixa Postal 109, CEP 79320-900 Corumbá, MS, Brazil. E-mail: sasantos@cpap.embrapa.br
3Statistician, Ph.D., Embrapa-Secretaria de Administração
Estratégica (SEA), CEP 70770-901 Brasília, DF, Brazil.
4Statistician, Dep. de Estatística, Universidade de Brasília,
CEP 70910-900 Brasília, DF, Brazil.
5Veterinarian, M.Sc., Embrapa-CPAP.
INTRODUCTION
Populations can be characterized by their growth
patterns but growth is a complex biological
phenom-enon with no adequately defined direct measure
tists, regardless of specialization, who are concerned
with the effects of their research and
recommenda-tions on lifetime production efficiency (Fitzhugh,
1976). Growth curves have been characterized by
plotting measures of size (weight, height, width)
against age. The use of regression models to describe
growth condenses the information contained in a
time series into a few biologically interpretable
pa-rameters (Eisen, 1974). Several models are
avail-able for this purpose and experimental comparisons
are needed to assist in the choice of the most
appro-priate model (Brown et al., 1976). Mathematical
nonlinear models to describe growth have been
de-veloped by Brody (1945), Richards (1959), Nelder
(1961), and Ratkowsky (1983). These models have
been used in beef cattle to describe weight-age
rela-tionships (Brown et al., 1972, 1976; Duarte, 1975;
De Nise & Brinks, 1985; Nobre et al., 1987) but
have not been used in horses.
The majority of the studies on horse growth are
based on weight, height, chest girth and cannon bone
(Green, 1961, 1969, 1976; Hintz et al., 1979; Santos,
1989). Of these measurements, height has more
prac-tical interest because it is required for the purpose
of description and classification of horses, while
weight is more a measure of the physical condition
of the animal (Reed & Dunn, 1977; Hickman &
Collis, 1984). According to Trenkle & Marple (1983)
growth curves can be of significant value in future
selection programs and in the study of the
energet-ics of growth.
The purpose of the present study was to choose,
among alternative nonlinear response functions, the
best model in regard to goodness of fit,
computa-tional ease, and validity of large sample
approxima-tions in small samples. The sex and age effects on
the growth of Pantaneiro horses in the Brazilian
Pantanal were also evaluated.
MATERIAL AND METHODS
The data on measurements of height at withers were obtained for 14 males and 12 females at Nhumirim farm, Nhecolândia sub-region, Brazilian Pantanal, during the period from 1990 to 1995. Monthly height at withers measures were taken until 3 months of age and at 6, 12, 18, 24, and 36 months of age. Eight animals were alsomeasured at 48 months of age. Most measurements were taken by the same person, however variation can occur due mainly to the levelness of the ground. Height at with-ers is a vertical distance from the ground to the highest protruding thoracic vertebrae. The measurements were taken with a hipometer. Only horses with 1080 or more days of age were included in the analyses. Breeding sea-sons lasted last four months (November to February) and three sires were used in this period. The animals were maintained in native pasture and they received common salt ad libitum. A growth curve was fit to each animal. Six nonlinear response functions were selected for curve fitting: 1. Brody growth function: Ht = a(1-be-kt); 2.
Richards growth function: Ht = a(1-be-kt)m; 3. Gompertz
growth function: Ht = a exp{-exp(b-kt)}; 4. Logistic
growth function: t b kt e 1 a H − +
= ; 5. Weibull growth func-tion: = − −ktδ t a be H ; 6. Morgan-Mercer-Flodin - MMF growth function: δδ + + = t k at bk
Ht . These models are
de-scribed in detail in Ratkowsky (1983) where one may also find handy techniques to obtain initial parameter estimates in each case. They are all used to fit the same type of data. The ordinary least squares estimates, for each animal and each model, were computed using the modified Gauss-Newton method of Hartley (1961) and PROC NLIN of SAS (1990). In regard to these fits five animals (four males and one female) showed atypical behavior and were ex-cluded from further analyses. In this context the number of animals under investigation was reduced to 21 (10 males and 11 females). For the excluded horses in all instances but one either convergence was not obtained or it was very slow for some of the models. In the exception case convergence results were sound but the parameter esti-mates were far different (outliers) suggesting the presence of a distinct population. The model MMF did not con-verge for any of the animals. This is the only instance were we can report a serious computational difficulty.
A common characteristic of all growth models above is that they share at least two biologically relevant param-eters. The asymptote a representing average size at matu-rity and the parameter k. The latter is a function of the ratio of the maximum growth rate to mature size. It is commonly referred as maturity index. For the Weibull model, the parameter δ δ δ δ δ can also be seen as a maturity
in-dex. The smaller and positive are its values the slower will be the approximation of the corresponding growth curves to their respective asymptotes. We attach no gen-eral special meaning to band to m, the latterin the case of the Richards response. They both influence the speed at which the asymptote is attained. With a positive associa-tion between band k, a large maturity index will imply a large b and the combined effect will be a fast maturity rate. Another parametric function of biological importance is the height at birth, which is given by a(1-b), a(1-b)m,
a exp {-eb}, a/(1+eb) and a-b, respectively, for models 1,
2, 3, 4, 5 e 6.
Two diagnostic measures were used to choose among the alternative response functions. A measure of nonlinearity (intrinsic curvature) and a measure of good-ness of fit. Curvature measures have been introduced in the statistical literature by Beale (1960) and extended by Bates & Watts (1980). The concept is extensively used by Ratkowsky (1983) to choose among alternative models and to assess the validity of large sample statistical proce-dures in small samples. It is a diagnostic tool intended to measure departures from linearity. The more linear a model is the more likely it is to show desirable computational and distributional properties for nonlinear least square estimates. Small curvatures imply that biases in the esti-mation of parameters and their variances will be negli-gible and the corresponding statistical distributions will be close to normal. Curvatures appear intwo kinds: in-trinsic curvature, a property of the response function in use, and parametric curvature, a property of the particular parametrization used in the definition of the response func-tion. A large parametric curvature may be reduced by a suitable reparametrization of the model function. A large intrinsic curvature, on the other hand, may put in jeop-ardy the use of asymptotic theory in a particular applica-tion. The concepts are presented in full detail in Seber & Wild (1989) were one may also find the computing algo-rithm suggested by Bates & Watts (1980). Oliveira (1996) developed a SAS macro to implement the Bates and Watts procedure which is used in this study to compute intrinsic curvatures.
The measure of goodness of fit used was the combined mean squares error (MSE).
For model i, i j N 1 j i j N 1 j i
df
SSE
MSE
= =∑
∑
=
where N is the number of animals,
SSE
ij is the residual sum of squares resulting from the fit of model i to the jth animal, anddf
ji is the corresponding error degrees of freedom.The presence of sex and age differences was investi-gated via t-tests on the parameters, exploring the replica-tion pattern imbedded in the experiment.
RESULTS AND DISCUSSION
Table 1 shows the average value of the intrinsic
curvature and the combined mean square error for
each of the nonlinear response functions considered
in this study. The two models with best prediction
power (smallest combined mean square error) are
Richards and Weibull. Between these, the Weibull
model showed a considerable lower level of
nonlinearity. It is worth to mention that the
coeffi-cients of determination for the Weibull fits range
from 0.988 to 0.999. The average value for females
is 0.994 and for males 0.993. Since the Richards
response function has an essentially equivalent
pre-diction power, similar results apply for the
corre-sponding model.
Table 2 shows average values of parameter
esti-mates for the Weibull model for each sex.
Confi-dence intervals were evaluated under the
assump-tion that parameter estimates represent samples from
normal populations centered at the corresponding
true population values. This approach is convenient
to deal with the correlation problem imbedded in
the study of growth curves. Although observations
taken on the same individual may be correlated,
pa-TABLE 1. Intrinsic curvature and combined mean square error for growth models according to sex.Model Sex Intrinsic
curvature mean squareCombined Brody Female 0.12619 4.05749 Male 0.11718 4.43629 Richards Female 0.58829 2.47668 Male 1.70630 3.02947 Gompertz Female 0.01045 5.18596 Male 0.01006 5.50570 Logistic Female 0.15674 5.50570 Male 0.14871 6.01068 Weibull Female 0.16394 2.50996 Male 0.18146 3.17833
Parameter Sex Average Lower confidence limit Upper confidence limit a Female 139.016 135.229 142.802 Male 140.225 136.830 143.621 b Female 58.913 52.534 65.292 Male 52.931 48.133 57.729 log(k) Female 0.372 0.245 0.498 Male 0.299 0.188 0.409 k Female 1.472 1.278 1.645 Male 1.360 1.207 1.505 δ Female 0.732 0.629 0.835 Male 0.835 0.776 0.894 a-b Female 80.103 74.068 86.137 Male 87.294 84.493 90.096
TABLE 2. Average values and 95% confidence intervals for parameters of the Weibull model.
TABLE 3. Values of t-tests for parameters of interest, Weibull model. The method of Satterthwaite was used for a-b (SAS, 1990). Parameter Df t-value p-value
a-b 14 -2.5668 0.0224
a 19 -0.5610 0.5813
b 19 1.7523 0.0958
δ 19 -2.0067 0.0592
log(k) 19 1.0244 0.3185
FIG. 1. Height model functions for males and females based on average values of parameters obtained with the Weibull model.
rameters, obtained for different individuals, inherit
the independence of the sampling scheme. Only the
maturity index
k
fails to pass Shapiro-Wilks
nor-mality test (SAS, 1990). However log(
k
) has a
dis-tribution that can be fairly closely approximated by
a normal. In this context it was computed a
confi-dence interval for log(
k
). The interval was inverted
to set up confidence bounds for
k
.
The asymptote
a
, the average height at maturity,
was only 1.2 cm higher for males than for females
(Table 2). The maturity index
k
was larger for
fe-males than fe-males (1.472 and 1.360, respectively).
The differences were not markedly significant since,
as shown in Table 3, the corresponding hypothesis
of equality cannot be rejected. The Weibull curves
for each sex, based on the average values of
Table 2, are shown in Fig. 1.
Only height at birth fails the test of homogeneity
of variances.
Table 3 reports the findings. Only a-b (height at
birth) showed a real sex difference. Females had a
lesser height at birth than males. This did not agree
with the findings of Reed & Dunn (1977) for the
Arabian horse. The parameters
δδδδδ
and
b
showed
mar-ginal significance. Females showed a smaller value
of
δδδδδ
and a larger value of
b
.
The relationship between asymptotic growth level
and maturity may be assessed from Table 4 where
Height (cm) 140 130 120 110 100 90 80 70 0 1 2 3 4 5 Age (years)
E(h) Males = 140.225 - 52.931 exp (-1.360t.835)
TABLE 4. Spearman (rank) correlation coefficients obtained by the Weibull model. Values in parenthesis are p-values for the null hypothesis that the coefficient is zero.
Parameter Male Female
a b δ k a b δ k a 1.00 0.75 -0.10 -0.07 1.00 0.44 -0.62 -0.57 (0.01) (0.78) (0.85) (0.18) (0.04) (0.07) b 1.00 -0.55 -0.08 1.00 -0.79 0.31 (0.09) (0.83) (0.00) (0.36) δ 1.00 0.12 1.00 -0.10 (0.75) (0.77) k 1.00 1.00
Spearman (rank) correlation coefficients are shown.
There is a negative association between
a
and
k
(and
δδδδδ
)
only for females.
The larger maturity index estimate and the
nega-tive association between
a
and
k
observed for the
female group indicate that females mature earlier
than males. This is confirmed by a smaller value of
δδδδδ
(and a larger
b
)
for the female group and the
sig-nificant negative association of this parameter with
the asymptote. A further indication the females
ma-ture earlier, following De Nise & Brinks (1985), is
the significant smaller height at birth for females and
the positive association between
b
and
k
for the
group. According to Brown et al. (1972) different
mature size may or may not represent different
pat-terns of growth. In the present study, the mature
height obtained for males was bellow the average
(142 cm ) obtained for Pantaneiro males registered
in the Association ABCCP. For females the
oppo-site was observed (the Association average is
137 cm ). Both values are within the confidence
lim-its of Table 3 but the value for males is quite close
to the upper bound. There are two explanations for
this. Firstly the males registered with the
associa-tion are selected animals and secondly the males
observed in the present study may not have reached
maturity. Reed & Dunn (1977) studying Arabian
horses observed that the mature height at withers
for females was achieved by 48 months of age
whereas males grew another 1.0 cm from 48 to 60
months of age. Therefore, in order to have a good fit
to asymptotic height, 60 months of age is necessary
for males. In this context is also important to
em-phasize that the animals we observed were
main-tained in native pastures without a feed supplement.
CONCLUSIONS
1. The Weibull response function is the best
para-metric model to fit height-age data for Pantaneiro
horses.
2. The male response curve slightly dominates.
ACKNOWLEDGEMENTS
To Maria Cristina Medeiros Mazza; for the
as-sistance to Roberto Rondon and Márcio Silva and
others; for the collaboration in the collection of the
data.
REFERENCES
BATES, D.M.; WATTS, D.G. Relative curvature mea-sures of nonlinearity. Journal of theRoyal Statis-tical Society, Series B, v.42, p.1-25, 1980. BEALE, E.M.L. Confidence regions in non-linear
esti-mation. Journal of the Royal Statistical Society, Series B, v.22, p.41-76, 1960.
BRODY, S. Bioenergetics and growth. New York: Reinhold, 1945. p.491-661.
BROWN, J.E.; BROWN, C.J.; BUTTS, W.T. A discus-sion of the genetic aspects of weight mature weight and rate of maturing in Hereford and Angus cattle.
Journal of AnimalScience, v.34, n.4, p.525-537, 1972.
BROWN, J.E.; FITZHUGH, H.; CARTWRIGHT, T.C. A comparison of nonlinear models for describing weight-age relationships in cattle. Journal of Ani-mal Science, v.42, n.4, p.810-818, 1976.
DE NISE, R.S.K.; BRINKS, J.S. Genetic and environ-mental aspects of the growth curve parameters in beef cows. Journal of Animal Science,v.61, n.6, p.1431-1440, 1985.
DUARTE, F.A.M. Estudo da curva de crescimento de animais de raça Nelore (Bos taurusindicus) através de cinco modelos estocásticos. Ribeirão Preto: USP, 1975. 284p. Tese de Mestrado. EISEN, E.J. Results of growth curve analyses in mice and
rats. Journal of Animal Science,v.42, n.4, p.1008-1023, 1974.
FITZHUGH, H. Analysis of growth curves and strategies for alterating their shape. Journal ofAnimal Sci-ence, v.42, n.4, p.1036-1051, 1976.
GREEN, D.A. A review of studies on the growth rate of the horse. British Veterinary Journal, v.117, p.181-191, 1961.
GREEN, D.A. A study of growth rate in thoroughbred foals. British Veterinary Journal, v.125, p.539-545, 1969.
GREEN, D.A. Growth rate in thoroughbred yearlings and two years olds. Equine VeterinaryJournal, v.8, n.3, p.133-134, 1976.
HARTLEY, H.O. The modified Gauss-Newton method for the fitting of nonlinear regression functions by least squares. Technometrics, v.3, p.269-280, 1961. HICKMAN, J.; COLLIS, C. Measurement of horses. The
Veterinary Record, v.114, p.491-493, 1984. HINTZ, H.F.; HINTZ, R.L.; VANVIECK, L.D. Growth
rate of thoroughbred. Effects of age of dam, year
and month of birth, and sex of foal. Journal of Animal Science, v.48, n.3, p.480-487, 1979. NELDER, J.A. The fitting of a generalization of the
lo-gistic curve. Biometrics, v.17, p.89-110, 1961. NOBRE, P.R.C.; ROSA, A.N.; SILVA, L.C.;
EVANGELISTA, S.R.M. Curvas de crescimento de gado Nelore ajustadas para diferentes frequências de pesagens. Pesquisa Agropecuária Brasileira,
Brasília,v.22, n.9/10, p.1027-1037, 1987. OLIVEIRA, M.R. Medidas de diagnóstico em regressão
não linear. Brasília: UnB, Departamento de Estatística, 1996. 35p. Relatório de Estágio Supervisionado.
RATKOWSKY, D.A. Nonlinear regression modeling: a unified practical approach. New York: Marcel Dekker, 1983. 276p.
REED, K.R.; DUNN, N.K. Growth and development of the Arabian horse. In: EQUINE NUTRITION AND PHYSIOLOGY SYMPOSIUM, 1977, St. Louis. Proceedings... Philadelphia: ENP Society, 1977. p.76-98.
RICHARDS, F.J. A flexible growth function for empiri-cal use. Journal of Experimental Biology, v.10, p.290-300, 1959.
ROBINSON, D.W. Growth patterns in swine. Journal of Animal Science, v.42, n.4, p.1024-1035, 1976. SANTOS, S.A. Estudo sobre algumas características
de crescimento de cavalos Brasileirode Hipismo produzidos no haras Pioneiro. Piracicaba: USP, 1989. 90p. Dissertação de Mestrado.
SAS INSTITUTE. SAS users guide: statistics,
version 6. 4.ed. Cary, NC, 1990. v.2, 1686p. SEBER, G.A.F.; WILD, C.J. Nonlinear regression. New
York: John Wiley, 1989. 768p.
TRENKLE, A.; MARPLE, D.N. Growth and development of meat animals. Journal of AnimalScience, v.57, n.2, p.273-283, 1983.